▶ YouTube Summaries

• https://www.youtube.com/watch?v=cVUVfn8OF5k
• Original title: Recovering from AI Psychosis | TheStandup

Adam (open-code maintainer, ex-vegan-meme target) joins TheStandup crew to recount a month-long burnout from "AI psychosis": running 4–5 Claude/Codex agents in parallel via git worktrees, prompting through the night, chasing slot-machine dopamine off occasional good outputs. He took five weeks fully offline (Netflix, golf, family), came back without the parallel-agent obsession, and reflects on triggers, signs, and how the team converged on a healthier outcome- driven workflow.

Timeline of the descent

• April 2025: started building open-code shortly after Claude Code shipped. Coded mostly by hand through summer; Opus 4 felt lazy and shortcut-prone, so Adam kept reviewing every line.
• December: GPT-5.2 era. Output noticeably more "clever" than Anthropic models. Adam switched to GPT, stopped reviewing as carefully, started parallelizing agents across git worktrees because runs took ~10 minutes.
• Dec–Jan: full obsession — 4–5 agents in flight, endless pointless refactors and performance audits, laptop in bedroom, waking up mid-night to fire off prompts. Lost sleep, couldn't go 30 minutes without a new prompt.
• Forced stop: ~5 weeks offline. First two weeks sick in bed, watched Breaking Bad. Then golf and tennis with wife. First real break in 15-year career.

GPT vs Claude in the "extremely online" crowd

• Adam and Anomaly team (20 devs) now default to GPT/Codex models.
• Claude reportedly preferred by people who don't read the code — $20 Anthropic plan yields ~$4,000 of tokens and dashboards nobody audits.
• Claude: lazy, shortcut-heavy, but better at UI. GPT: slower, more thoughtful, weirdly insists on auto-generating tech-stack headers ("Hello, this is an app built with effect and bun") in UI output.
• DHH-style fast open-source models (Kimmy K2.6) suit prompt-and-go workflows; bigger models better for scoping whole projects.
• Public perception of Anthropic-lead is largely vibes; recent developer-relations backlash has shifted sentiment to OpenAI.

What AI psychosis looks like

• Multiple Codex/Claude subscriptions, multiple accounts, agents running overnight, "polyphasic sleep so I can prompt more" (Andreessen clip referenced).
• Convincing yourself you're more productive when really chasing the slot-machine reward of a good completion.
• Stops reading diffs. Adam later found 500 duplicate is_record functions in his codebase — had to add eslint rules banning model- favored function names.
• Measuring success in PR/diff size rather than outcomes.
• Gary Tan's USB-C-cable-burned-out tweet cited as canonical late-stage symptom (crew can't tell if satire).

Signs you're slipping

• Sleep loss is the leading indicator (Adam's wife flagged it first).
• Abandoning prior healthy habits (gym, hobbies) for "one more prompt."
• Agreeing with everything a model lab CEO says when they have billions on the line to mislead you.
• Hyping a product you've never used ("zero is awesome, can't wait to try it").
• "CC Ryan Flurry — delete your account" reply as informal diagnostic.

Detox and re-entry

• Five weeks no LLMs, minimal X/Twitter. X identified as a major driver — constant competitor visibility.
• Back ~1.5 weeks at time of recording. No parallel agents now; faster GPT models removed the latency that made parallelism feel necessary. Reviewing diffs again. Catching up in Discord between prompts instead of firing another agent.
• Team-wide shift: focus on outcome quality, not churn volume.
• Still nervous about relapse — doesn't know exactly what triggered it the first time (probably "this model is smarter than the last" feeling around GPT-5.2).

Joke pitch that became serious

• Build an opt-in "AI psychosis detector" into open-code: warn when N agents have run for >24h, optional friend-voted activation ("9 friends flagged you").

Advice for solo devs without an intervening spouse

• Audit basics: sleep, exercise, social contact. Same as any addiction.
• Treat the surrounding online culture as adversarial — for every friend telling you to slow down, ten X accounts will say go harder.
• Likely has to run its course; intervention is hard when the whole peer group is mid-psychosis too.
• https://www.youtube.com/watch?v=9tcLy9TnPDU
• Original title: Github's Wildest Hack Yet

Prime walks through the May 19, 2026 GitHub breach: ~3,800 internal repos exfiltrated via a poisoned NX Console VS Code extension on a compromised employee device. Group "PCP" listed source + ~4,000 repos on Breached Forums threatening a free leak. Broader takeaway: VS Code extensions are an NPM-backed supply chain with no Microsoft endorsement, published binaries needn't match GitHub source, and auto-update silently pulls poisoned versions.

Timeline

• 2026-05-19 — GitHub discloses unauthorized access, claims no customer data impact.
• ~24h later — PCP posts on Breached Forums advertising source + ~4,000 private repos, free leak if no buyer, explicitly not ransom.
• Shortly after — GitHub confirms vector: compromised employee device via poisoned VS Code extension. ~3,800 repos count consistent with investigation.

Root cause

Malicious NX Console (Nx DevTools) extension, confirmed by Nx's CEO. Supply-chain compromise pushed the poisoned version through VS Code auto-update.

Key points

• VS Code attack surface is huge: NPM-backed, anyone can publish, no Microsoft endorsement, published artifact doesn't have to match GitHub source.
• Auto-update is the core vulnerability — disable it, audit extensions.
• VS Code opens whole project trees including .env secrets, while compromised extensions run continuously against them.
• NPM remains an ongoing breach vector (cf. recent "Shai Hulud" attacks); editor-extension ecosystems on NPM inherit the risk.
• Owning GitHub's source expands the attack surface against GitHub itself; expect faster follow-on exploits.
• Prime says he's wary of storing sensitive private repos / GitHub Secrets there, on top of cited 86.68% uptime over 90 days (81 incidents).
• Closing joke: "Should have used Vim."